AI Potluck
Contribute
Back to Gap Map Model components / Language-specific datasets

Rasa

AI4Bharat
gated / Overall score: 3.0

Rasa is an expressive text-to-speech corpus covering the 22 officially recognized languages of India, with a female and a male voice per language and about 1,145 hours in total. Its 48 kHz recordings mix neutral Wikipedia readings, speech in the six Ekman emotions, voice-assistant style commands, conversation, news reading and book narration. AI4Bharat built it, and the method appeared at Interspeech 2024.

The card cites an Interspeech 2024 paper but links no arXiv entry, so none is listed.

Openness

3 high confidence
3.0
license
cc-by-4.0(card License section and metadata)
access
auto(Hugging Face gate asks for contact information and approves automatically)
dataset_card
present(card describes speaking styles and per-speaker hours)

Every recording is under an attribution-only license, released through a Hugging Face form that approves automatically once contact details are given.

Adoption

3 high confidence
3.0

Hugging Face downloads of the single corpus repository through the automatic gate. Use through voices trained on it, such as IndicF5, is not counted.

Capability

3 medium confidence
3.0

Rasa records a female and a male studio voice in each of India's twenty-two official languages across emotions and speaking styles, and speech synthesis models from AI Bharat and SPRING Lab are trained on it. At about eleven hundred hours from a few dozen speakers it is small beside English speech corpora of hundreds of thousands of hours, and IndicVoices from the same team records tens of thousands of speakers.

  • https://huggingface.co/datasets/ai4bharat/Rasa recorded 2026-09-24

    "the first high-quality multilingual expressive Text-to-Speech (TTS) dataset for any Indian language"; "44 speaker-language pairs across all 22 Indian languages"; models trained include ai4bharat/IndicF5, SPRINGLab/Indic-Mio

Verified 2026-09-24